Deep Learning Framework for Multiclass Detection of Ocular Diseases in Fundus Images

The early and accurate detection of eye diseases is paramount in preventing irreversible vision loss and facilitating timely treatment. Conventional diagnostic strategies generally depend on subjective analysis of experts leading to variability in diagnosis. Convolutional Neural Networks (CNNs) have evolved as a prospective solution to classify diseases using medical images with remarkable accuracy. This study proposes a CNN-based methodology for the multiclass classification of ocular diseases, including diabetic retinopathy (DR), cataract, and glaucoma. The objective is to improve the detection and evaluation of these conditions, thereby enabling effective intervention and patient management. The research introduces a 9-layer CNN designed for the automated classification of eye disorders, utilizing two datasets of fundus images. The CNN proficiently distinguishes between normal and disease-related features. To enhance the model's performance, preprocessing techniques and hyperparameter optimization are applied. The model is implemented using TensorFlow and Python within a Jupyter Notebook environment. With a learning rate (LR) set at 0.0001 and a batch size (BS) of 8, the proposed CNN achieves a training accuracy of 99.94% and a testing accuracy of 89.82% on the first dataset. When the batch size is increased to 32 while keeping the learning rate at 0.0001, the CNN model attains a training accuracy of 99.97% and a testing accuracy of 96.15% on the second dataset. The results indicate that this deep learning (DL) model demonstrates outstanding performance in classifying DR, cataract, glaucoma, and healthy eye conditions from fundus images, and the proposed approach can assist ophthalmologists in accurately diagnosing eye diseases.